The same attacks under a control protocol. Protocols spend usefulness to buy safety — deferring to a trusted model, resampling suspicious actions, or editing them before they run.